Quantifying the sensitivity of seismic facies classification to seismic attribute selection: An explainable machine learning study

2022 ◽  
pp. 1-90
Author(s):  
David Lubo-Robles ◽  
Deepak Devegowda ◽  
Vikram Jayaram ◽  
Heather Bedle ◽  
Kurt J. Marfurt ◽  
...  
Keyword(s):  
Machine Learning ◽  
Input Data ◽  
Global Analysis ◽  
Random Noise ◽  
Seismic Attributes ◽  
Attribute Selection ◽  
Seismic Facies ◽  
Attribute Importance ◽  
Seismic Attribute ◽  
Facies Classification

During the past two decades, geoscientists have used machine learning to produce a more quantitative reservoir characterization and to discover hidden patterns in their data. However, as the complexity of these models increase, the sensitivity of their results to the choice of the input data becomes more challenging. Measuring how the model uses the input data to perform either a classification or regression task provides an understanding of the data-to-geology relationships which indicates how confident we are in the prediction. To provide such insight, the ML community has developed Local Interpretable Model-agnostic Explanations (LIME), and SHapley Additive exPlanations (SHAP) tools. In this study, we train a random forest architecture using a suite of seismic attributes as input to differentiate between mass transport deposits (MTDs), salt, and conformal siliciclastic sediments in a Gulf of Mexico dataset. We apply SHAP to understand how the model uses the input seismic attributes to identify target seismic facies and examine in what manner variations in the input such as adding band-limited random noise or applying a Kuwahara filter impact the models’ predictions. During our global analysis, we find that the attribute importance is dynamic, and changes based on the quality of the seismic attributes and the seismic facies analyzed. For our data volume and target facies, attributes measuring changes in dip and energy show the largest importance for all cases in our sensitivity analysis. We note that to discriminate between the seismic facies, the ML architecture learns a “set of rules” in multi-attribute space and that overlap between MTDs, salt, and conformal sediments might exist based on the seismic attribute analyzed. Finally, using SHAP at a voxel-scale, we understand why certain areas of interest were misclassified by the algorithm and perform an in-context interpretation to analyze how changes in the geology impact the model’s predictions.

Predicting and 3D modeling of karst zones using seismic facies analysis in Ordovician carbonates of the Tahe oilfield, China

2020 ◽  
Vol 8 (2) ◽  
pp. T293-T307
Author(s):  
José N. Méndez ◽  
Qiang Jin ◽  
María González ◽  
Wei Hehua ◽  
Cyril D. Boateng
Keyword(s):  
Cross Sections ◽  
Input Data ◽  
Optimal Number ◽  
Seismic Facies ◽  
Rock Properties ◽  
Seismic Attribute ◽  
Clustering Technique ◽  
Yingshan Formation ◽  
Facies Classification ◽  
Seismic Facies Analysis

Karsted carbonates of the Ordovician Yingshan Formation represent significant hydrocarbon reservoirs in the Tarim Basin, China. Due to the geologic complexity of the formation, realistically predicting and modeling karst zones and rock properties is challenging. This drives the need to apply diverse techniques for building a suitable geologic model. We have developed a static model approach that uses fully automated seismic facies classification processes for predicting and modeling patterns associated with karst elements. Our method uses a seismic attribute and well logs as input data. We initially processed a seismic facies volume using the hierarchical clustering technique. This is based on seismic attribute values that take into account an optimal number of classes. The outcome reveals various patterns illustrated with low amplitudes highlighting the geomorphology of paleokarst elements. Simultaneously, a seismic traces map of the karsted interval was processed using the hybrid clustering technique conducted on seismic trace shape. In this case, the karst facies was extracted from the output and used as secondary input data in trend analysis of the model. Both outputs obtained from clustering techniques are processed in a volume of the most probable facies, which delineate the karst patterns. The results of the modeling process are visualized in various time slices and cross sections, appropriately recognizing the relationship of estimated patterns with karst zones. We have evaluated the karstification thickness and porosity map obtained from the 3D model that detail a reasonable connectivity between karst elements. This is based on the paleogeographic location and type of filling, as well as the dissolution development along the main striking faults. Finally, our method outputs a logical model of karst zones located within the host rock, which reduces the uncertainty and identify nonperforated segments.

Characterizing a turbidite system in Canterbury Basin, New Zealand, using seismic attributes and distance-preserving self-organizing maps

2016 ◽  
Vol 4 (1) ◽  
pp. SB79-SB89 ◽  
Author(s):  
Tao Zhao ◽  
Jing Zhang ◽  
Fangyu Li ◽  
Kurt J. Marfurt
Keyword(s):  
New Zealand ◽  
Seismic Attributes ◽  
Seismic Facies ◽  
Seismic Attribute ◽  
Self Organizing Maps ◽  
Seismic Amplitude ◽  
Facies Classification ◽  
Turbidite System ◽  
Reliable Classification ◽  
Self Organizing

Recent developments in seismic attributes and seismic facies classification techniques have greatly enhanced the capability of interpreters to delineate and characterize features that are not prominent in conventional 3D seismic amplitude volumes. The use of appropriate seismic attributes that quantify the characteristics of different geologic facies can accelerate and partially automate the interpretation process. Self-organizing maps (SOMs) are a popular seismic facies classification tool that extract similar patterns embedded with multiple seismic attribute volumes. By preserving the distance in the input data space into the SOM latent space, the internal relation among data vectors on an SOM facies map is better presented, resulting in a more reliable classification. We have determined the effectiveness of the modified algorithm by applying it to a turbidite system in Canterbury Basin, offshore New Zealand. By incorporating seismic attributes and distance-preserving SOM classification, we were able to observe architectural elements that are overlooked when using a conventional seismic amplitude volume for interpretation.

Independent component analysis for reservoir geomorphology and unsupervised seismic facies classification in the Taranaki Basin, New Zealand

2019 ◽  
Vol 7 (3) ◽  
pp. SE19-SE42 ◽  
Author(s):  
David Lubo-Robles ◽  
Kurt J. Marfurt
Keyword(s):  
New Zealand ◽  
Independent Component Analysis ◽  
Order Statistics ◽  
Input Data ◽  
Component Analysis ◽  
Independent Component ◽  
Seismic Facies ◽  
Blue Color ◽  
Independent Components ◽  
Facies Classification

During the past two decades, the number of volumetric seismic attributes has increased to the point at which interpreters are overwhelmed and cannot analyze all of the information that is available. Principal component analysis (PCA) is one of the best-known multivariate analysis techniques that decompose the input data into second-order statistics by maximizing the variance, thus obtaining mathematically uncorrelated components. Unfortunately, projecting the information in the multiple input data volumes onto an orthogonal basis often mixes rather than separates geologic features of interest. To address this issue, we have implemented and evaluated a relatively new unsupervised multiattribute analysis technique called independent component analysis (ICA), which is based on higher order statistics. We evaluate our algorithm to study the internal architecture of turbiditic channel complexes present in the Moki A sands Formation, Taranaki Basin, New Zealand. We input 12 spectral magnitude components ranging from 25 to 80 Hz into the ICA algorithm and we plot 3 of the resulting independent components against a red-green-blue color scheme to generate a single volume in which the colored independent components correspond to different seismic facies. The results obtained using ICA proved to be superior to those obtained using PCA. Specifically, ICA provides improved resolution and separates geologic features from noise. Moreover, with ICA, we can geologically analyze the different seismic facies and relate them to sand- and mud-prone seismic facies associated with axial and off-axis deposition and cut-and-fill architectures.

Genetic‐algorithm/neural‐network approach to seismic attribute selection for well‐log prediction

Geophysics ◽  
2004 ◽  
Vol 69 (1) ◽  
pp. 212-221 ◽  
Author(s):  
Kevin P. Dorrington ◽  
Curtis A. Link
Keyword(s):  
Neural Network ◽  
Genetic Algorithm ◽  
Reservoir Characterization ◽  
Seismic Attributes ◽  
Optimal Number ◽  
Attribute Selection ◽  
Well Log ◽  
Seismic Attribute ◽  
Neural Network Approach ◽  
Data Set

Neural‐network prediction of well‐log data using seismic attributes is an important reservoir characterization technique because it allows extrapolation of log properties throughout a seismic volume. The strength of neural‐networks in the area of pattern recognition is key in its success for delineating the complex nonlinear relationship between seismic attributes and log properties. We have found that good neural‐network generalization of well‐log properties can be accomplished using a small number of seismic attributes. This study presents a new method for seismic attribute selection using a genetic‐algorithm approach. The genetic algorithm attribute selection uses neural‐network training results to choose the optimal number and type of seismic attributes for porosity prediction. We apply the genetic‐algorithm attribute‐selection method to the C38 reservoir in the Stratton field 3D seismic data set. Eleven wells with porosity logs are used to train a neural network using genetic‐algorithm selected‐attribute combinations. A histogram of 50 genetic‐algorithm attribute selection runs indicates that amplitude‐based attributes are the best porosity predictors for this data set. On average, the genetic algorithm selected four attributes for optimal porosity log prediction, although the number of attributes chosen ranged from one to nine. A predicted porosity volume was generated using the best genetic‐algorithm attribute combination based on an average cross‐validation correlation coefficient. This volume suggested a network of channel sands within the C38 reservoir.

scholarly journals Value of Geologically Derived Features in Machine Learning Facies Classification

2019 ◽  
Vol 52 (1) ◽  
pp. 5-29 ◽  
Author(s):  
Julie Halotel ◽  
Vasily Demyanov ◽  
Andy Gardiner
Keyword(s):  
Machine Learning ◽  
Expert Knowledge ◽  
Random Noise ◽  
Uncertainty Assessment ◽  
Support Vector ◽  
Volume Data ◽  
Facies Classification ◽  
Geological Features ◽  
Output Noise ◽  
Very Large Datasets

AbstractThe aim of this work is to demonstrate how geologically interpretative features can improve machine learning facies classification with uncertainty assessment. Manual interpretation of lithofacies from wireline log data is traditionally performed by an expert, can be subject to biases, and is substantially laborious and time consuming for very large datasets. Characterizing the interpretational uncertainty in facies classification is quite difficult, but it can be very important for reservoir development decisions. Thus, automation of the facies classification process using machine learning is a potentially intuitive and efficient way to facilitate facies interpretation based on large-volume data. It can also enable more adequate quantification of the uncertainty in facies classification by ensuring that possible alternative lithological scenarios are not overlooked. An improvement of the performance of purely data-driven classifiers by integrating geological features and expert knowledge as additional inputs is proposed herein, with the aim of equipping the classifier with more geological insight and gaining interpretability by making it more explanatory. Support vector machine and random forest classifiers are compared to demonstrate the superiority of the latter. This study contrasts facies classification using only conventional wireline log inputs and using additional geological features. In the first experiment, geological rule-based constraints were implemented as an additional derived and constructed input. These account for key geological features that a petrophysics or geological expert would attribute to typical and identifiable wireline log responses. In the second experiment, geological interpretative features (i.e., grain size, pore size, and argillaceous content) were used as additional independent inputs to enhance the prediction accuracy and geological consistency of the classification outcomes. Input and output noise injection experiments demonstrated the robustness of the results towards systematic and random noise in the data. The aspiration of this study is to establish geological characteristics and knowledge to be considered as decisive data when used in machine learning facies classification.

A seismic facies classification method based on the convolutional neural network and the probabilistic framework for seismic attributes and spatial classification

2019 ◽  
Vol 7 (3) ◽  
pp. SE225-SE236 ◽  
Author(s):  
Zhege Liu ◽  
Junxing Cao ◽  
Yujia Lu ◽  
Shuna Chen ◽  
Jianli Liu
Keyword(s):  
Neural Network ◽  
Convolutional Neural Network ◽  
Early Stage ◽  
Seismic Attributes ◽  
Seismic Facies ◽  
Classification Method ◽  
Probabilistic Framework ◽  
Spatial Classification ◽  
Facies Classification ◽  
Maximal Information Coefficient

In the early stage of oil and gas exploration, due to the lack of available drilling data, the automatic seismic facies classification technology mainly relies on the unsupervised clustering method combined with the seismic multiattribute. However, the clustering results are unstable and have no clear geologic significance. The supervised classification method based on manual interpretation can provide corresponding geologic significance, but there are still some problems such as the discrete classification results and low accuracy. To solve these problems, inspired by hyperspectral and spatial probability distribution technology, we have developed a classification framework called the probabilistic framework for seismic attributes and spatial classification (PFSSC). It can improve the continuity of the classification results by combining the classification probability and the spatial partial probability of the classifier output. In addition, the convolutional neural network (CNN) is a typical classification algorithm in deep learning. By convolution and pooling, we could use it to extract features of complex nonlinear objects for classification. Taking advantage of the combination of PFSSC and CNN, we could effectively solve the existing problems mentioned above in seismic facies classification. It is worth mentioning that we select seismic the multiattribute by maximal information coefficient (MIC) before the seismic facies classification. Finally, using the CNN-PFSSC and MIC methods, we can obtain high accuracy in the test set, reasonable continuity within the same seismic facies, clear boundaries between different seismic facies, and seismic facies classification results consistent with sedimentological laws.

scholarly journals Seismic facies analysis using machine learning

Geophysics ◽  
2018 ◽  
Vol 83 (5) ◽  
pp. O83-O95 ◽  
Author(s):  
Thilo Wrona ◽  
Indranil Pan ◽  
Robert L. Gawthorpe ◽  
Haakon Fossen
Keyword(s):  
Machine Learning ◽  
Seismic Facies ◽  
Machine Learning Techniques ◽  
Seismic Section ◽  
Seismic Reflection Data ◽  
Reflection Data ◽  
Facies Classification ◽  
Learning Techniques ◽  
Seismic Facies Analysis ◽  
Northern North Sea

Seismic interpretations are, by definition, subjective and often require significant time and expertise from the interpreter. We are convinced that machine-learning techniques can help address these problems by performing seismic facies analyses in a rigorous, repeatable way. For this purpose, we use state-of-the-art 3D broadband seismic reflection data of the northern North Sea. Our workflow includes five basic steps. First, we extract seismic attributes to highlight features in the data. Second, we perform a manual seismic facies classification on 10,000 examples. Third, we use some of these examples to train a range of models to predict seismic facies. Fourth, we analyze the performance of these models on the remaining examples. Fifth, we select the “best” model (i.e., highest accuracy) and apply it to a seismic section. As such, we highlight that machine-learning techniques can increase the efficiency of seismic facies analyses.

Enhanced ant tracking: Using a multispectral seismic attribute workflow to improve 3D fault detection

2021 ◽  
Vol 40 (7) ◽  
pp. 502-512
Author(s):  
Mateo Acuña-Uribe ◽  
María Camila Pico-Forero ◽  
Paul Goyes-Peñafiel ◽  
Darwin Mateus
Keyword(s):  
Random Noise ◽  
Seismic Attributes ◽  
Hydrocarbon Exploration ◽  
Fault Identification ◽  
Complex Task ◽  
Frequency Filtering ◽  
Seismic Attribute ◽  
Interpretation Process ◽  
Seismic Volumes ◽  
Fault Interpretation

Fault interpretation is a complex task that requires time and effort on behalf of the interpreter. Moreover, it plays a key role during subsurface structural characterization either for hydrocarbon exploration and development or well planning and placement. Seismic attributes are tools that help interpreters identify subsurface characteristics that cannot be observed clearly. Unfortunately, indiscriminate and random seismic attribute use affects the fault interpretation process. We have developed a multispectral seismic attribute workflow composed of dip-azimuth extraction, structural filtering, frequency filtering, detection of amplitude discontinuities, enhancement of amplitude discontinuities, and automatic fault extraction. The result is an enhanced ant-tracking volume in which faults are improved compared to common fault-enhanced workflows that incorporate the ant-tracking algorithm. To prove the effectiveness of the enhanced ant-tracking volume, we have applied this methodology in three seismic volumes with different random noise content and seismic characteristics. The detected and extracted faults are continuous, clean, and accurate. The proposed fault identification workflow reduces the effort and time spent in fault interpretation as a result of the integration and appropriate use of various types of seismic attributes, spectral decomposition, and swarm intelligence.

Sign in / Sign up

Export Citation Format

Share Document